Reconstruction error fault detection method based on graph neural network

Abstract

The present invention provides a reconstruction error fault detection method based on a graph neural network. The method obtains time series data of industrial equipment, wherein the time series data includes features of multiple nodes; processes the time series data using a convolutional neural network to obtain the temporal features of each node; constructs a graph structure based on the time series data, wherein the graph structure represents the spatial associations between nodes; processes the graph structure and the temporal features using a graph neural network to obtain the spatial features of each node; reconstructs the spatial features using a decoder to obtain a reconstructed output; and determines a fault detection result based on the reconstructed output and the time series data. This method improves the accuracy and robustness of fault detection. This method is suitable for real-time monitoring and predictive maintenance of complex industrial systems and has important practical value.

Description

Technical Field

[0001] The present invention relates to the field of neural networks, and in particular to a reconstruction error fault detection method based on graph neural networks. Background Art

[0002] In the large hydropower sector, the continuity and stability of equipment operation are crucial to the safety of the power system. However, the acquisition of actual fault case data is often very limited, especially for the operating data of key equipment. Due to the low frequency of faults and the difficulty of large-scale collection, it is difficult to obtain sufficient labeled data to directly train fault prediction models. Traditional methods mainly rely on reconstruction algorithms based on historical data. These methods detect anomalies by modeling and reconstructing the operating status of equipment. However, such methods have shortcomings in extracting complex temporal and spatial correlation features. In particular, when there is a high correlation between different monitoring points, the reconstruction algorithm has difficulty capturing the complex relationship between multiple measuring points, resulting in the accuracy and stability of the prediction effect being difficult to meet actual needs.

[0003] Existing technologies for predicting time-series data from industrial equipment generally face the problem of insufficient spatial feature modeling. This is particularly true in industrial systems with complex connections between multiple measurement points. Traditional reconstruction algorithms struggle to capture the spatial dependencies between devices, resulting in insufficient prediction accuracy. The application of graph neural networks (GNNs) to industrial equipment, leveraging their efficient processing capabilities for graph-structured data, can comprehensively capture the spatial dependencies between devices. In large hydropower systems, GNNs can effectively model the complex relationships between different sensor or device nodes, explore the linkage characteristics between multiple measurement points, and construct more accurate spatial dependency models that adapt to the nonlinearity, heterogeneity, and complex network structures of devices. Summary of the Invention

[0004] The main purpose of the present invention is to provide a reconstruction error fault detection method based on graph neural network to solve the problem of insufficient spatial feature modeling commonly faced in the existing technology in the prediction of industrial equipment time series data.

[0005] To solve the above technical problems, the technical solution adopted by the present invention is: a reconstruction error fault detection method based on graph neural network, the method comprising:

[0006] Obtain the time series data of industrial equipment, which includes the characteristics of multiple nodes. Assume that the input time series data is , where B is the batch size, W is the time window length, N is the number of nodes, and each node has multiple features;

[0007] Processing the time series data through a convolutional neural network to obtain the time characteristics of each node;

[0008] Constructing a graph structure based on the time series data, wherein the graph structure represents the spatial relationship between nodes;

[0009] Processing the graph structure and the temporal features through a graph neural network to obtain spatial features of each node;

[0010] Reconstructing the spatial features through a decoder to obtain a reconstructed output;

[0011] A fault detection result is determined according to the reconstructed output and the time series data.

[0012] In a preferred embodiment, the time series data is processed by a convolutional neural network to obtain a time feature of each node, including: performing linear projection on the time series data to obtain a projection feature;

[0013] Performing convolution processing on the projection feature through a first layer of one-dimensional convolution to obtain a first convolution output;

[0014] Normalizing and activating the first convolution output to obtain a first time feature;

[0015] The first time feature is convolved by a second layer of one-dimensional convolution, and a residual connection is added to obtain the time feature.

[0016] In a preferred embodiment, linear projection is performed on the time series data to obtain projection features;

[0017] The formula is: , the projection feature formula is: ;

[0018] Where X represents the input industrial equipment time series data, It is the projection feature obtained by linear projection of X, and its dimension is , is the weight matrix in the linear projection operation, is the bias vector in the linear projection operation, Indicates the number of output channels of the convolutional neural network;

[0019] Performing convolution processing on the projection feature through a first layer of one-dimensional convolution to obtain a first convolution output;

[0020] The first convolution output is normalized and activated to obtain the first time feature. Assuming the convolution kernel size of the first layer CNN is 1, the output is ,in , after normalization and ReLU activation function processing;

[0021] The first time feature is convolved by the second layer of one-dimensional convolution and a residual connection is added to obtain the time feature. The formula is: , the final feature The temporal structure information of the input is preserved and mapped to the convolutional feature dimension.

[0022] In a preferred embodiment, the graph structure is constructed based on the time series data, and the graph structure represents the spatial relationship between nodes, including:

[0023] For each time step in the time series data, calculate the neighborhood distance of each node feature in a preset space;

[0024] Determine the adjacency relationship based on the preset radius threshold, generate an adjacency matrix, and generate an edge set based on the adjacency matrix , the edge set is used to construct the graph structure, the formula is:

[0025] ;

[0026] generating an edge set according to the adjacency matrix, wherein the edge set is used to construct the graph structure;

[0027] in, It represents the distance between the feature vectors of node i and node j, r is the preset radius threshold, which is a pre-set constant.

[0028] In a preferred embodiment, the graph structure and the temporal features are processed by a graph neural network to obtain the spatial features of each node, including:

[0029] Inputting the temporal features and the graph structure into a graph attention convolutional layer;

[0030] The first-layer spatial features are obtained by aggregating the adjacent node features in the graph structure through the attention mechanism. The formula is: ;

[0031] Normalize and perform residual connection processing on the first layer of spatial features to obtain intermediate features;

[0032] Repeat the graph attention convolution process to obtain the spatial features, repeat the three-layer GATv2 convolution, add residual connection and normalization to each layer output, and obtain the final spatial aggregated node feature representation .

[0033] In a preferred embodiment, the spatial features are reconstructed by a decoder to obtain a reconstructed output, including:

[0034] Inputting the spatial features into a multi-layer perceptron decoder;

[0035] Performing dimension transformation on the spatial features through a fully connected layer to obtain decoding features;

[0036] Linearly map the decoded features to generate a reconstructed output consistent with the dimension of the time series data. The formula is: , the final output .

[0037] In a preferred embodiment, determining a fault detection result based on the reconstructed output and the time series data includes:

[0038] Calculating a reconstruction error between the reconstructed output and the time series data;

[0039] If the reconstruction error exceeds a preset threshold, it is determined that the corresponding node has a fault;

[0040] The overall fault status of the industrial equipment is determined based on the fault judgment results of each node.

[0041] In a preferred embodiment, the time series data of the industrial equipment is obtained, and the time series data includes characteristics of multiple nodes, including:

[0042] Acquire raw data from multiple sensors in industrial equipment;

[0043] Preprocessing the raw data to obtain time series data including a time window and the number of nodes;

[0044] Feature extraction is performed on the time series data to obtain a feature vector for each node.

[0045] In the preferred embodiment, the method comprehensively models the temporal variation trend and spatial correlation of measurement point data by combining convolutional neural networks and graph neural networks to improve the accuracy of fault warning.

[0046] In the preferred solution, the spatial relationship between measurement points is automatically constructed through the radius-based dynamic neighborhood graph structure generation method to adapt to different equipment and measurement point configurations, and the robustness of the algorithm on noisy data and nonlinear data is enhanced through layer-by-layer feature extraction and residual fusion of multi-layer CNN and GATv2.

[0047] The present invention provides a reconstruction error fault detection method based on graph neural network, which realizes accurate fault identification by integrating time series data processing and graph structure analysis. The method first obtains the time series data of multiple nodes of the equipment, uses convolutional neural network to extract time features, and constructs a graph structure based on the spatial correlation between nodes. Then, the graph neural network is used to process the graph structure and time features to obtain the node spatial features. The decoder is then used to reconstruct the spatial features and generate a reconstructed output. Finally, the fault is judged by comparing the error between the reconstructed output and the original data. The present invention innovatively combines time series and spatial information, uses deep learning methods to automatically extract features, avoids manual feature engineering, can effectively capture equipment status anomalies, and improves the accuracy and robustness of fault detection. This method is suitable for real-time monitoring and predictive maintenance of complex industrial systems and has important practical value. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] The present invention will be further described below with reference to the accompanying drawings and examples:

[0049] Figure 1 It is a flow chart of the reconstruction error fault detection method based on graph neural network of the present invention; DETAILED DESCRIPTION

[0050] Example 1

[0051] like Figure 1 As shown, a reconstruction error fault detection method based on graph neural network includes:

[0052] Obtain the time series data of industrial equipment, which includes the characteristics of multiple nodes. Assume that the input time series data is , where B is the batch size, W is the time window length, N is the number of nodes, and each node has features;

[0053] Processing the time series data through a convolutional neural network to obtain the time characteristics of each node;

[0054] Constructing a graph structure based on the time series data, wherein the graph structure represents the spatial relationship between nodes;

[0055] Processing the graph structure and the temporal features through a graph neural network to obtain spatial features of each node;

[0056] Reconstructing the spatial features through a decoder to obtain a reconstructed output;

[0057] A fault detection result is determined according to the reconstructed output and the time series data.

[0058] In a preferred embodiment, the time series data is processed by a convolutional neural network to obtain a time feature of each node, including: performing linear projection on the time series data to obtain a projection feature;

[0059] Performing convolution processing on the projection feature through a first layer of one-dimensional convolution to obtain a first convolution output;

[0060] Normalizing and activating the first convolution output to obtain a first time feature;

[0061] The first time feature is convolved by a second layer of one-dimensional convolution, and a residual connection is added to obtain the time feature.

[0062] In a preferred embodiment, linear projection is performed on the time series data to obtain projection features;

[0063] The formula is: , the projection feature formula is: ;

[0064] Where X represents the input industrial equipment time series data, It is the projection feature obtained by linear projection of X, and its dimension is , is the weight matrix in the linear projection operation, is the bias vector in the linear projection operation, Indicates the number of output channels of the convolutional neural network;

[0065] Performing convolution processing on the projection feature through a first layer of one-dimensional convolution to obtain a first convolution output;

[0066] The first convolution output is normalized and activated to obtain the first time feature. Assuming the convolution kernel size of the first layer CNN is 1, the output is ,in , after normalization and ReLU activation function processing;

[0067] The first time feature is convolved by the second layer of one-dimensional convolution and a residual connection is added to obtain the time feature. The formula is: , the final feature The temporal structure information of the input is preserved and mapped to the convolutional feature dimension.

[0068] In a preferred embodiment, the graph structure is constructed based on the time series data, and the graph structure represents the spatial relationship between nodes, including:

[0069] For each time step in the time series data, calculate the neighborhood distance of each node feature in a preset space;

[0070] Determine the adjacency relationship based on the preset radius threshold, generate an adjacency matrix, and generate an edge set based on the adjacency matrix , the edge set is used to construct the graph structure, the formula is:

[0071] ;

[0072] generating an edge set according to the adjacency matrix, wherein the edge set is used to construct the graph structure;

[0073] in, It represents the distance between the feature vectors of node i and node j, r is the preset radius threshold, which is a pre-set constant.

[0074] In a preferred embodiment, the graph structure and the temporal features are processed by a graph neural network to obtain the spatial features of each node, including:

[0075] Inputting the temporal features and the graph structure into a graph attention convolutional layer;

[0076] The first-layer spatial features are obtained by aggregating the adjacent node features in the graph structure through the attention mechanism. The formula is: ;

[0077] Normalize and perform residual connection processing on the first layer of spatial features to obtain intermediate features;

[0078] Repeat the graph attention convolution process to obtain the spatial features, repeat the three-layer GATv2 convolution, add residual connection and normalization to each layer output, and obtain the final spatial aggregated node feature representation .

[0079] In a preferred embodiment, the spatial features are reconstructed by a decoder to obtain a reconstructed output, including:

[0080] Inputting the spatial features into a multi-layer perceptron decoder;

[0081] Performing dimension transformation on the spatial features through a fully connected layer to obtain decoding features;

[0082] Linearly map the decoded features to generate a reconstructed output consistent with the dimension of the time series data. The formula is: , the final output .

[0083] In a preferred embodiment, determining a fault detection result based on the reconstructed output and the time series data includes:

[0084] Calculating a reconstruction error between the reconstructed output and the time series data;

[0085] If the reconstruction error exceeds a preset threshold, it is determined that the corresponding node has a fault;

[0086] The overall fault status of the industrial equipment is determined based on the fault judgment results of each node.

[0087] In a preferred embodiment, the time series data of the industrial equipment is obtained, and the time series data includes characteristics of multiple nodes, including:

[0088] Acquire raw data from multiple sensors in industrial equipment;

[0089] Preprocessing the raw data to obtain time series data including a time window and the number of nodes;

[0090] Feature extraction is performed on the time series data to obtain a feature vector for each node.

[0091] In the preferred embodiment, the method comprehensively models the temporal variation trend and spatial correlation of measurement point data by combining convolutional neural networks and graph neural networks to improve the accuracy of fault warning.

[0092] In the preferred solution, the spatial relationship between measurement points is automatically constructed through the radius-based dynamic neighborhood graph structure generation method to adapt to different equipment and measurement point configurations, and the robustness of the algorithm on noisy data and nonlinear data is enhanced through layer-by-layer feature extraction and residual fusion of multi-layer CNN and GATv2.

[0093] Example 2

[0094] Further illustrate with reference to Example 1, Figure 1 As shown in the structure, the algorithm aims to extract local information of industrial equipment node features through convolutional neural networks (CNN) and model the spatial association between nodes through graph neural networks (GNN) to improve the accuracy of equipment status prediction.

[0095] (1) Preparation and initial projection of input data

[0096] Assume that the input time series data is , where B is the batch size, W is the time window length, N is the number of nodes, and each node has features (e.g. 5).

[0097] Initial projection: First, the input features are mapped to the same dimension as the CNN output channels through a linear layer. , which enables CNN to extract local information more effectively:

[0098] ;

[0099] in, ;

[0100] (2) Temporal feature extraction and normalization

[0101] 1) One-dimensional convolution: Use a 1D convolution layer to convolve the projected features to extract temporal features. Assuming the convolution kernel size of the first layer of CNN is 1, the output is:

[0102]

[0103] in, , after normalization and ReLU activation function processing.

[0104] 2) Second convolution layer: Based on the output of the first convolution layer, 1D convolution and residual connection are applied again to further extract temporal features:

[0105]

[0106] The final features The temporal structure information of the input is preserved and mapped to the convolutional feature dimension.

[0107] (3) Graph structure construction and spatial feature extraction

[0108] In order to model the spatial relationship between nodes at multiple measurement points, the nodes are constructed into a graph structure using a radius neighborhood graph.

[0109] 1) Edge set generation: For each time step, the neighborhood of the node feature in the Euclidean space is calculated, and the adjacency matrix is ​​generated with the radius r as the threshold to form the edge set edge_index.

[0110] ;

[0111] 2) GATv2 convolutional layer: Based on the adjacency matrix, in each layer of graph convolution, the features of adjacent nodes are spatially aggregated, and the attention mechanism is used to learn the weights of important nodes:

[0112]

[0113] Repeat three layers of GATv2 convolution, add residual connection and normalization to the output of each layer, and finally represent the node features of spatial aggregation

[0114] (4) Multi-layer decoding and reconstruction

[0115] The features processed by the convolutional layer and graph convolutional layer are input into the multi-layer perceptron (MLP) decoder to generate the final reconstructed output.

[0116] Decoding layer: Through the fully connected layer decoder, the features of each node are reconstructed into the original input dimension:

[0117]

[0118] Final Output , used to reconstruct the node status.

[0119] This technical solution combines CNN with GATv2 to extract local features in the time dimension and capture the correlation information between nodes in the spatial dimension through graph neural network, realizing the reconstruction and prediction of multi-node time series data in complex industrial systems.

[0120] Traditional time series prediction algorithms typically only extract single features in time or space. This makes it difficult to accurately capture the correlations between measurement points in complex equipment systems with multiple measurement points, making it difficult to achieve high-precision early warnings. By combining convolutional neural networks (CNNs) and graph neural networks (GNNs), this algorithm can comprehensively model the temporal trends and spatial correlations of measurement point data, making the early warning model more accurate in identifying potential anomalies and significantly improving the accuracy of fault warnings.

[0121] The algorithm automatically constructs spatial relationships between measurement points through the generation of a dynamic, radius-based neighborhood graph structure. Compared to traditional methods that rely on fixed topology, the adaptive graph structure capture mechanism of this invention can generate accurate spatial structure models based on different equipment and measurement point configurations. This innovative design makes the early warning system more adaptable to complex inter-device correlation patterns, improving the adaptability of the early warning model to different equipment layouts, thereby enhancing the accuracy of early warnings in practical applications.

[0122] By using a multi-layer CNN and GATv2 to extract features and fuse residuals layer by layer, the algorithm is more robust on noisy and nonlinear data, eliminating interference from irrelevant factors to a certain extent and accurately capturing early signs of equipment failure. This feature makes the early warning model more stable, maintaining high prediction accuracy even when measurement point data fluctuates or anomalies occur at industrial sites.

[0123] Example 3

[0124] Further illustrating data acquisition and preprocessing with reference to Example 1: Raw data is acquired from various sensors within the chemical equipment cluster, including temperature, pressure, and flow rate sensors. These sensors are distributed across different devices and collect data at varying sampling rates, such as every 10 seconds for temperature sensors and every 5 seconds for pressure sensors. This raw data is transmitted to a data processing center via industrial Ethernet, where data format conversion is performed during transmission to ensure data compatibility.

[0125] The raw data is preprocessed to solve the time alignment problem first. Linear interpolation is used to unify data with different sampling frequencies into the same timestamp sequence.

[0126] Time feature extraction: Assume that the input time series data is , where the batch size B is 32, the time window length W is 200, the number of nodes N is 50, and each node has 5 features.

[0127] Perform linear projection on the time series data and map the original 5-dimensional features to 32 dimensions compatible with subsequent convolution operations through a fully connected layer. The formula is: , and get the projection feature .

[0128] The projection features are convolved using the first layer of one-dimensional convolution, with the convolution kernel size set to 3 and the number of output channels to 64. After the convolution operation, the first convolution output is obtained, and the formula is , the output dimension is The first convolution output is batch normalized and ReLU activated to stabilize the training process and enhance feature expression capabilities.

[0129] The first time feature is convolved through the second layer of one-dimensional convolution. The convolution kernel size is set to 5, the number of output channels is 128, and a residual connection is added. The formula is ,The obtained temporal features retain the temporal structure information of the input and enhance the feature expression capability.

[0130] Graph structure construction: For each time step in the preprocessed time series data, the neighborhood distance of each node feature in the preset space is calculated. In the chemical equipment scenario, the reactor is used as a node, and the Euclidean distance is used to measure the spatial relationship between nodes. Assume that at a certain moment, the coordinates of reactor A are (x1, y1, z1) and the coordinates of reactor B are (x2, y2, z2). Calculate the distance between the two.

[0131] The adjacency relationship is determined based on the preset radius threshold (set to 50 meters) to generate an adjacency matrix. If the distance between the two reactors is less than the threshold, the corresponding position in the adjacency matrix is ​​set to 1, otherwise it is set to 0. The formula is .

[0132] The edge set is generated based on the adjacency matrix and used to construct the graph structure. At the same time, in order to more accurately reflect the strength of the association between nodes, the edge weight is set to the inverse of the distance.

[0133] Spatial feature extraction: The temporal features and the constructed graph structure are input into the Graph Attention Convolutional Layer (GATv2). Before input, the temporal features are aligned to the same dimension as the graph node features through linear projection (assuming the graph node feature dimension is 256, and the temporal features are mapped from 128 dimensions to 256 dimensions through a fully connected layer).

[0134] GATv2 uses an attention mechanism to aggregate features of adjacent nodes in the graph. It computes a nonlinear transformation of query-key vector pairs. For example, for node E and its neighbors F and G, the model calculates the attention scores between E and F, and between E and G, respectively. It then introduces nonlinearity using the LeakyReLU activation function and obtains weights after Softmax normalization. Assuming the scores are 1.5 and 0.8, respectively, the normalized weights are 0.68 and 0.32. The first layer of spatial features after aggregation consists of a weighted combination of 60% of node E's own features and 40% of its neighbors' features.

[0135] The first layer of spatial features undergoes layer normalization and residual connections. Layer normalization scales the mean and variance based on the feature dimension, while residual connections add the normalized results to the original input to mitigate the vanishing gradient problem. Graph attention convolutions are repeated, and the second layer, GATv2, further aggregates high-order neighbor information to ultimately obtain spatial features for each node.

[0136] Reconstruction and Fault Detection: The resulting spatial features are fed into a multi-layer perceptron decoder. The multi-layer perceptron consists of multiple fully connected layers that progressively extract and combine features using nonlinear activation functions. After multiple layers of processing, the 256-dimensional spatial features are mapped to 64-dimensional decoded features through fully connected layers.

[0137] Perform linear mapping on the decoded features to generate a reconstructed output consistent with the dimensions of the original time series data. Assuming that each node of the original time series data is 5-dimensional, the 64-dimensional decoded features are mapped to 5 dimensions through linear mapping. The formula is , and get the reconstructed output.

[0138] The reconstruction error between the reconstructed output and the original time series data is calculated, using the mean square error (MSE) as the calculation metric. If the reconstruction error exceeds a preset threshold (set to 0.6 through statistical analysis of historical normal data), the corresponding node is deemed faulty. Based on the fault diagnosis results for each node, the overall fault status of the chemical equipment cluster is determined.

[0139] Example 4

[0140] In conjunction with Example 1, it is further described that time series data of industrial equipment is obtained, and the time series data includes features of multiple nodes; the time series data is processed by a convolutional neural network to obtain the time features of each node.

[0141] When obtaining time series data from industrial equipment, it is usually necessary to collect data from multiple sensors or measurement points. This data may include various characteristics such as temperature, pressure, vibration, and current.

[0142] In a wind turbine, multiple sensors are installed on each blade, gearbox, and generator. These sensors record device status information at regular intervals (e.g., once per second). This data constitutes multi-node time series data, and the characteristics of each node may include temperature, vibration frequency, rotational speed, etc. Collecting this data provides a foundation for subsequent analysis and modeling. When processing time series data using convolutional neural networks, the convolutional layers can capture local patterns and trends in the data.

[0143] For temperature data from a particular node, a convolutional neural network can use a sliding window approach to extract local features from the time series, such as rising or falling temperature trends and periodic fluctuations. For example, if the temperature data for a particular node has shown a gradual upward trend over the past 10 seconds, the convolutional neural network can use the convolution kernel to extract this trend and use it as the temporal feature of the node. This processing method effectively captures short-term changes in the data, providing important information for subsequent predictions. Convolutional neural networks typically use one-dimensional convolution kernels when processing time series data because time series data is one-dimensional.

[0144] For vibration data from a particular node, a one-dimensional convolution kernel can be slid across the time dimension to extract local features from the vibration signal, such as peaks, valleys, or vibration patterns at specific frequencies. For example, if a node's vibration data has exhibited multiple peaks over the past five seconds, the convolutional neural network can use the convolution kernel to extract the locations and intensities of these peaks and use them as temporal features for that node. This processing method effectively captures key events in the data, supporting subsequent anomaly detection. After the convolutional neural network extracts temporal features, these features can be used for further analysis and modeling.

[0145] In wind turbine fault warnings, extracted temporal features can be used to determine whether the equipment is in an abnormal state. For example, if the temperature signature of a node shows a sharp increase in a short period of time, this may indicate a risk of overheating. By combining the temporal features of other nodes, a more comprehensive assessment of the equipment's condition can be achieved, thereby improving the accuracy of fault warnings. This processing approach can effectively improve the accuracy of industrial equipment status predictions and provide strong support for equipment maintenance and management.

[0146] A graph structure is constructed based on the time series data, where the graph structure represents the spatial association between nodes; the graph structure and the time features are processed through a graph neural network to obtain the spatial features of each node.

[0147] The key to building a graph structure lies in defining adjacency relationships between nodes. For example, in a network of temperature sensors for industrial equipment, an edge connection is established if the physical distance between two sensors is less than a preset radius threshold (e.g., 5 meters). Suppose a workshop has 10 temperature sensors deployed. The pairwise distances are calculated using 3D coordinates. When the distance between sensor A (coordinates 1, 2, 0) and sensor B (coordinates 4, 2, 0) is 3 meters, the threshold condition is met, and an edge AB is generated. This physical distance-based graph construction directly reflects the spatial coupling between devices. For example, sensors near a high-temperature heat source will form a strong correlation due to heat conduction. Graph neural networks use a message passing mechanism to aggregate neighborhood information.

[0148] Specifically, each node fuses neighboring features using an aggregation function (such as mean pooling). For example, suppose the initial feature of node C is [75°C, 0.8MPa], and the features of its neighboring nodes D and E are [72°C, 0.7MPa] and [78°C, 0.9MPa], respectively. After aggregation, the intermediate feature of node C is generated as [75°C, 0.8MPa]⊕[72°C, 0.7MPa]⊕[78°C, 0.9MPa], where ⊕ represents feature concatenation or weighted summation. This operation can capture local temperature distribution patterns. For example, when the temperatures of most of its neighbors rise abnormally, the feature of node C will imply a "heat diffusion" trend. The fusion of temporal and spatial features is achieved through a gating mechanism. The node temporal features output by the LSTM (e.g., [0.3, -0.1], representing the recent temperature change rate) are concatenated with the spatial features generated by the GNN (e.g., [0.4, 0.2], representing the spatial gradient). The concatenation is then mapped into a unified representation using a fully connected layer. For example, the temporal characteristics of a vibration sensor node reflect a "continuous increase in amplitude," while the spatial characteristics indicate a "decreasing phase difference between adjacent nodes." After fusion, this fusion can infer a composite state of "forward propagation of mechanical vibration waves." This multi-dimensional fusion can improve fault location accuracy. For example, when spatially related nodes simultaneously exhibit high-frequency vibrations, the fault source can be determined to be located in the center of the equipment cluster. Sparsification of the adjacency matrix improves computational efficiency. For a system with 100 nodes, if the radius threshold is set to 10 meters, each node is connected to only 3-5 neighbors on average, and the proportion of non-zero elements in the adjacency matrix is ​​less than 5%. This sparsity allows GNN training to process only valid connections. For example, with GPU acceleration, the computational complexity of a single graph convolution is reduced from O(N²) to O(kN), where k is the average number of neighbors. In practical applications, a wind turbine gearbox monitoring system has reduced graph inference time from 200ms to 40ms using this method, meeting real-time diagnostic requirements. The introduction of edge attributes enhances relationship modeling. In addition to being connected, edge weights can also reflect the strength of spatial associations. For example, the edge weight between two pressure sensors is set to 1 / (1 + distance). When nodes F and G are 2 meters apart, the edge weight is 0.33, while the edge weight with node H, which is 5 meters away, is 0.17. During GNN aggregation, closer neighbors are given higher weights. For example, in the feature update of node F, G's contribution is twice that of H, consistent with the distance decay effect in the physical world. Using this method in a hydraulic system, the leak location error was reduced from 1.5 meters to 0.6 meters.

[0149] The spatial features are reconstructed by a decoder to obtain a reconstructed output; and a fault detection result is determined according to the reconstructed output and the time series data.

[0150] The decoder reconstructs the spatial features to produce a reconstructed output. The core of this process is to map the high-dimensional features extracted through graph convolution and temporal features back to the original input dimensions. The decoder typically uses a multi-layer perceptron (MLL) architecture, where each layer progressively reduces the high-dimensional features using nonlinear activation functions, ultimately outputting a reconstructed result with the same dimensionality as the original input data.

[0151] In industrial equipment fault detection scenarios, the original input may be multidimensional sensor data from the equipment. After graph convolution and time series feature extraction, the decoder reconstructs these features into the same dimensionality as the original sensor data, enabling intuitive comparison between the reconstructed data and the real data. This reconstruction process not only preserves the spatial and temporal characteristics of the data but also enables the identification of potential anomalies by comparing the reconstructed output with the real data. The key to determining the fault detection result based on the reconstructed output and the time series data is to identify anomalous patterns by comparing the differences between the reconstructed data and the real data.

[0152] In a power system, the reconstructed output may be a predicted value of power load, while the true data is the actual measured power load. By calculating the error between the two, it is possible to determine whether an anomaly exists. If the error exceeds a preset threshold, it may indicate a fault in the power system. The advantage of this method is that it can simultaneously utilize spatial and temporal characteristics to improve the accuracy of fault detection.

[0153] In wind turbine fault detection, the reconstructed output reflects the generator's operating status, while the time series data records the generator's actual operating parameters. By comparing the two, abnormal vibration or temperature rise can be detected promptly, providing early warning and preventing equipment damage. The technical benefit of this fault detection method lies in its ability to combine spatial and temporal features to improve detection sensitivity and accuracy.

[0154] In rail transit systems, the reconstructed output reflects the train's operating status, while the time series data records the train's actual operating parameters. By comparing the two, abnormal vibrations or speed fluctuations can be detected promptly, providing early warnings and preventing accidents. This approach is also adaptable to complex industrial environments. For example, in a chemical plant, the reconstructed output reflects the equipment's operating status, while the time series data records its actual operating parameters. By comparing the two, abnormal pressure or temperature changes can be detected, providing early warnings and preventing production accidents. The widespread application of this approach not only improves the safety of industrial equipment but also reduces maintenance costs, resulting in significant economic benefits.

[0155] The time series data is processed by a convolutional neural network to obtain the time features of each node, including: linearly projecting the time series data to obtain projection features; convolving the projection features through a first layer of one-dimensional convolution to obtain a first convolution output; normalizing and activating the first convolution output to obtain a first time feature; convolving the first time feature through a second layer of one-dimensional convolution, and adding a residual connection to obtain the time feature.

[0156] Linear projection is performed on the time series data to obtain projection features. The purpose of linear projection is to map the feature dimensions of the original time series data to dimensions compatible with subsequent convolution operations, so as to more efficiently extract temporal features.

[0157] Assuming the feature dimension of the original time series data is 10, and the subsequent convolutional layer expects an input dimension of 32, linear projection is used to map the 10-dimensional features to 32 dimensions. This step can be implemented using a fully connected layer with a weight matrix of 10×32 dimensions. Linear projection not only adjusts the feature dimension but also reduces noise and enhances features to a certain extent, providing higher-quality feature input for subsequent convolution operations. The projected features are then convolved with the first layer of one-dimensional convolution to produce the first convolution output. One-dimensional convolution is suitable for time series data because it can capture local patterns in the time series.

[0158] Assuming the projected feature dimension is 32 and the time step is 100, a one-dimensional convolutional layer with a kernel size of 3 and 64 output channels is used for processing. The convolution operation slides over each time step, extracting local temporal features and ultimately outputting a feature matrix of 64 × 100 dimensions. This step effectively extracts short-term dependencies in time series data, such as capturing short-term price fluctuations in stock price forecasting. The first convolution output is normalized and activated to obtain the first temporal feature. Normalization typically uses batch normalization or layer normalization to stabilize the training process and accelerate convergence.

[0159] Batch normalization is performed on the first convolutional output, which has a dimension of 64×100, so that the feature distribution of each channel has a mean of 0 and a variance of 1. Activation processing typically uses the ReLU function, which sets negative values ​​to 0 and retains positive values. This step can enhance the expressiveness of features while avoiding the vanishing gradient problem.

[0160] In speech recognition tasks, normalization and activation can make features more robust and reduce the impact of noise on the model. The first temporal features are convolved with a second layer of one-dimensional convolution and a residual connection is added to obtain temporal features. The kernel size and number of output channels of the second layer of one-dimensional convolution can be adjusted according to task requirements. For example, a convolution layer with a kernel size of 5 and 128 output channels can be used. The residual connection adds the output of the first and second convolutions, allowing the model to retain shallow features while learning deeper features.

[0161] In action recognition tasks, residual connections can simultaneously capture short-term action patterns and long-term action trends, improving the model's recognition accuracy. The resulting temporal features not only retain the temporal structure of the input, but also enhance the expressive power of the features through multi-layer convolution and residual connections.

[0162] The graph structure is constructed based on the time series data, and the graph structure represents the spatial association between nodes, including: for each time step in the time series data, calculating the neighborhood distance of each node feature in a preset space; determining the adjacency relationship based on a preset radius threshold and generating an adjacency matrix; generating an edge set based on the adjacency matrix, and the edge set is used to construct the graph structure.

[0163] When calculating neighborhood distance, Euclidean distance can be used to measure the spatial relationship between nodes. For example, in a wind farm's SCADA system, each wind turbine is considered a node, and its latitude and longitude coordinates serve as spatial features. For data at time step t=1, the spherical distance between node A (120.5°E, 32.8°N) and node B (120.6°E, 32.9°N) is calculated to be 15.2 kilometers, corrected for the curvature of the Earth. This spatial distance calculation accurately reflects the actual geographic location of the devices, laying the foundation for subsequent adjacency determination. The radius threshold setting should be tailored to the physical characteristics of the specific application scenario. For example, in oil pipeline monitoring, if the radius threshold is set to 50 meters, pressure sensor node C and flow meter node D, which are 48 meters apart, will be considered adjacent, while a connection will not be established with temperature sensor node E, which is 52 meters away. The threshold setting typically takes into account the sensor signal propagation range or the physical connection distance between devices. Too large a threshold will result in invalid connections, while too small a threshold will miss key connections. The adjacency matrix generation process reflects dynamic topological characteristics. In an urban water supply network, if the distance between pressure nodes F and G is 80 meters (less than the threshold of 100 meters) during a certain period, the corresponding position in the adjacency matrix is ​​set to 1. When pipeline pressure changes cause the hydraulic distance between the two nodes to increase to 110 meters, the corresponding position in the matrix is ​​reset to zero at the next time step. This dynamic adjustment can adapt to changes in system state and better reflect actual operating conditions than a static topology. Edge set construction must address sparse connectivity. In a substation equipment network, 500 nodes may only generate 3,000 edges (with an average degree of 6). In this case, storing the adjacency matrix in compressed sparse row format can save 90% of memory. Edge weights can be set to the inverse of distance. For example, the edge weight between transformer node H and circuit breaker node J, which are 20 meters apart, is 0.05 to reflect the spatial attenuation effect. The final integration of the graph structure requires the integration of multi-source features. For a chemical plant equipment graph, in addition to spatial adjacency edges, pipeline connection edges from the process flow diagram can be overlaid to form a multidimensional association network. For example, the reactor node K establishes edges with the spatially adjacent feed pump (25 meters away) and the process-connected discharge valve. This composite graph structure can more comprehensively represent system associations.

[0164] The graph structure and the temporal features are processed by a graph neural network to obtain spatial features of each node, including: inputting the temporal features and the graph structure into a graph attention convolution layer; aggregating adjacent node features in the graph structure through an attention mechanism to obtain a first layer of spatial features; normalizing and performing residual connection processing on the first layer of spatial features to obtain intermediate features; and repeating the graph attention convolution processing to obtain the spatial features.

[0165] When inputting temporal features and graph structures into the graph attention convolutional layer, the time series data must first be aligned to the same dimension as the graph node features through linear projection.

[0166] If the original temporal feature dimension is 128 and the graph node feature dimension is 256, the 128 dimension is mapped to 256 dimensions through a fully connected layer to ensure compatibility with subsequent convolution operations. The projected features and the adjacency matrix are input into the GATv2 layer together, at which point the attention mechanism calculates the dynamic weights between nodes. Taking traffic flow prediction as an example, a certain intersection node will dynamically assign attention weights to the intersections connected to it based on temporal features such as traffic speed and direction. For example, the weight of the upstream intersection may be increased to 0.8 during the morning rush hour and reduced to 0.2 at night. When aggregating adjacent node features through the attention mechanism, GATv2 first calculates a nonlinear transformation of the query-key vector pair.

[0167] For node A and its neighbors B and C, the model calculates attention scores for A and B, and A and C, respectively, and then introduces nonlinearity using the LeakyReLU activation function. Assuming the scores are 1.2 and 0.5, respectively, the weights after Softmax normalization are 0.71 and 0.29. The aggregated first-layer spatial features may contain a weighted combination of 70% of node A's own features and 30% of its neighbors' features. For example, if A's traffic feature is [0.6, 0.4] and B's is [0.8, 0.2], the aggregated result is [0.66, 0.34]. When normalizing and performing residual connections on the first-layer spatial features, layer normalization (LayerNorm) scales the mean and variance based on the feature dimension. For example, if a node feature has a mean of 0.5 and a variance of 1.2, normalization may adjust it to a standard distribution with a mean of 0 and a variance of 1. Residual connections add the normalized result to the original input. For example, if the original input is [0.6, 0.4], the normalized output is [-0.1, 0.1], and the final intermediate feature is [0.5, 0.5]. This operation can mitigate vanishing gradients, for example, by preserving key information from the initial temporal features in deep networks. When repeating the graph attention convolution process, the second layer of GATv2 further aggregates high-order neighbor information. For example, if node A's neighbor B is also connected to node D, the second layer may indirectly pass D's features to A with a weight of 0.3. If the third layer incorporates dynamic edge weights (such as weather effects), the spatial features can incorporate multi-order neighborhoods and external factors. The final output spatial features may include a fusion of three layers of information: the node's own temporal patterns (such as periodic traffic), local neighborhood status (such as congestion propagation), and global dependencies (such as the impact of regional events).

[0168] The spatial features are reconstructed by a decoder to obtain a reconstructed output, including: inputting the spatial features into a multi-layer perceptron decoder; performing dimensionality transformation on the spatial features through a fully connected layer to obtain a decoding feature; and linearly mapping the decoding feature to generate a reconstructed output consistent with the dimension of the time series data.

[0169] The spatial features are fed into the multilayer perceptron decoder to gradually extract and combine features through multiple layers of nonlinear transformations, thereby better capturing complex patterns in the data. The multilayer perceptron consists of multiple fully connected layers, each of which introduces nonlinearity through activation functions, enabling the model to learn more advanced feature representations.

[0170] When processing time series data, spatial features may contain state information from multiple nodes. Multi-layer perceptrons (MLPs) enable deep integration of this state information to generate more representative features. The fully connected layer transforms the spatial features to map them to a space consistent with the target output dimensions. Each neuron in a fully connected layer is connected to all neurons in the previous layer, enabling a global linear combination of the input features.

[0171] Assuming the spatial feature dimension is 128 and the target output dimension is 64, the fully connected layer can map the 128-dimensional features to 64 dimensions through the weight matrix, providing suitable input for the subsequent linear mapping. Linear mapping of the decoded features generates a reconstructed output consistent with the time series data dimension. This is done to restore the decoded features to the format of the original data. Linear mapping adjusts the decoded features to the same dimension as the original time series data through a simple matrix multiplication.

[0172] If the time series data has a dimension of 10, linear mapping can map the 64-dimensional decoded features to 10 dimensions, generating the final reconstructed output. This step ensures that the model accurately restores the original data while retaining the key information extracted from the spatial features. Through the above steps, the model can effectively convert spatial features into reconstructed output consistent with the original time series data. The multi-layer perceptron decoder provides powerful feature extraction capabilities, the fully connected layers enable flexible dimensionality transformation, and the linear mapping ensures output accuracy and consistency. This design not only improves the model's expressiveness but also enhances its adaptability and robustness in complex data scenarios.

[0173] The fault detection result is determined based on the reconstructed output and the time series data, including: calculating the reconstruction error between the reconstructed output and the time series data; if the reconstruction error exceeds a preset threshold, it is determined that a fault exists in the corresponding node; and based on the fault judgment result of each node, the overall fault status of the industrial equipment is determined.

[0174] When calculating the reconstruction error between the reconstructed output and the time series data, it's important to first understand that the reconstructed output is generated through a multi-layer decoding and reconstruction process, with the goal of restoring the characteristics of the original time series data as closely as possible. Reconstruction error is typically calculated using metrics such as mean squared error (MSE) or mean absolute error (MAE) to measure the difference between the reconstructed output and the original time series data.

[0175] In industrial equipment monitoring, if the original time series data for a node is [1.0, 2.0, 3.0] and the reconstructed output is [1.1, 2.2, 3.1], the reconstruction error can be calculated by calculating the sum of squares of the differences at each time point and taking the average. This error calculation intuitively reflects the accuracy of the reconstruction process and provides a basis for subsequent fault diagnosis. If the reconstruction error exceeds a preset threshold, the corresponding node is considered faulty. The threshold is determined based on the specific application scenario and historical data, and is typically determined through statistical analysis or experimental verification.

[0176] For a particular piece of industrial equipment, historical data shows that the reconstruction error for healthy nodes is typically less than 0.5, so a threshold of 0.5 can be set. When the reconstruction error for a node reaches 0.8, it is determined that the node is faulty. This threshold-based judgment method is simple and efficient, enabling rapid identification of abnormal nodes and providing timely warnings for equipment maintenance. When determining the overall fault status of the industrial equipment based on the fault judgment results for each node, it is necessary to comprehensively consider the fault conditions of all nodes.

[0177] Consider an industrial device consisting of 10 nodes. If the reconstruction error of three of these nodes exceeds a threshold, the overall status of the device can be determined based on the number and distribution of faulty nodes. If the faulty nodes are concentrated in a key module, this may indicate a serious problem with that module; if the faulty nodes are more dispersed, this may indicate a decline in overall device performance. This comprehensive analysis helps operations and maintenance personnel more accurately locate problems and develop targeted maintenance strategies. Through the above steps, the fault detection method based on reconstruction error can effectively identify abnormal nodes in industrial equipment and determine the overall fault status of the device. This method not only improves fault detection accuracy but also reduces false alarm rates, effectively ensuring efficient equipment operation.

[0178] The method of obtaining time series data of industrial equipment, wherein the time series data includes features of multiple nodes, includes: obtaining raw data from multiple sensors in the industrial equipment; preprocessing the raw data to obtain time series data including a time window and a number of nodes; and extracting features from the time series data to obtain a feature vector for each node.

[0179] Acquiring raw data from multiple sensors in industrial equipment typically involves physical signal acquisition and digital conversion. For example, in a hydropower plant monitoring scenario, vibration sensors collect shaft radial displacement signals at a 10kHz sampling rate, temperature sensors record bearing temperature every 5 seconds, and current transformers synchronously capture the three-phase winding current waveforms. These heterogeneous sensors convert analog signals into digital signals using industrial bus protocols (such as Modbus TCP), forming a multi-source data stream containing timestamps, device IDs, and raw measurement values. Due to varying sensor accuracy, current signals may retain three decimal places, while temperature data only requires one decimal place. This heterogeneity requires data type adaptability in subsequent processing modules. The key to raw data preprocessing lies in addressing data fragmentation and time alignment. For example, in a wind turbine fleet, the gearbox oil temperature sensor uploads data every minute, while the power transmitter outputs data at a sub-second frequency. Linear interpolation is required to pad the low-frequency data to a consistent timestamp sequence. When dividing windows, if the analysis objective is gear fault warning, a 300-second sliding window is required to cover the entire vibration characteristic cycle. If the analysis objective is power prediction, a 60-second window is more suitable for grid scheduling requirements. A dynamic thresholding method is used for outlier processing. For example, if the wind speed sensor readings exceed the historical mean ±3σ for three consecutive cycles, they are automatically replaced with the sliding median. Normalization is performed based on node type: temperature data is scaled to [0, 1] using the Min-Max method, while vibration acceleration is normalized using the Z-score. Feature extraction from time series data requires a balance between local details and global patterns. For motor current signals, a 1D CNN can capture short-term fluctuations: a convolution kernel with a width of 5 is used to scan the original waveform and extract a 5-dimensional feature vector, such as harmonic distortion rate and crest factor, within each time window. For pressure sensor data, frequency domain energy distribution features are extracted after STFT transformation. In the spatial dimension, if a graph structure is constructed from 20 temperature measurement points in a chemical reactor, GATv2 calculates the attention weights of adjacent nodes. For example, the weight coefficient of node A to node B is 0.83, reflecting the strength of the heat conduction path. The feature fusion stage combines the temporal CNN output with the spatial GAT features to form a final 128-dimensional node representation. This combined representation improves prediction accuracy by 12.7% compared to single-feature prediction.

[0180] The above embodiments are merely preferred technical solutions of the present invention and should not be construed as limiting the present invention. The scope of protection of the present invention shall be the technical solutions set forth in the claims, including equivalent alternatives to the technical features of the technical solutions set forth in the claims. In other words, equivalent alternatives and improvements within this scope are also within the scope of protection of the present invention.

Claims

1. A reconstruction error fault detection method based on graph neural network, characterized by: The method includes: Obtain the time series data of industrial equipment, which includes the characteristics of multiple nodes. The input time series data is , where B is the batch size, W is the time window length, N is the number of nodes, and each node has multiple features; Processing the time series data through a convolutional neural network to obtain the time characteristics of each node; Constructing a graph structure based on the time series data, wherein the graph structure represents the spatial relationship between nodes; Processing the graph structure and the temporal features through a graph neural network to obtain spatial features of each node; Reconstructing the spatial features through a decoder to obtain a reconstructed output; determining a fault detection result according to the reconstructed output and the time series data; The graph structure and the temporal features are processed by a graph neural network to obtain spatial features of each node, including: Inputting the temporal features and the graph structure into a graph attention convolutional layer; The first-layer spatial features are obtained by aggregating the adjacent node features in the graph structure through the attention mechanism. The formula is: ; Normalize and perform residual connection processing on the first layer of spatial features to obtain intermediate features; Repeat the graph attention convolution process to obtain the spatial features, repeat the three-layer GATv2 convolution, add residual connection and normalization to each layer output, and obtain the final spatial aggregated node feature representation ; The spatial features are reconstructed by a decoder to obtain a reconstructed output, including: Inputting the spatial features into a multi-layer perceptron decoder; Performing dimension transformation on the spatial features through a fully connected layer to obtain decoding features; Linearly map the decoded features to generate a reconstructed output consistent with the dimension of the time series data. The formula is: , the final output .

2. The method for detecting reconstruction error faults based on graph neural networks according to claim 1, wherein: The time series data is processed by a convolutional neural network to obtain the time characteristics of each node, including: Performing linear projection on the time series data to obtain projection features; Performing convolution processing on the projection feature through a first layer of one-dimensional convolution to obtain a first convolution output; Normalizing and activating the first convolution output to obtain a first time feature; The first time feature is convolved by a second layer of one-dimensional convolution, and a residual connection is added to obtain the time feature.

3. The method for detecting reconstruction error faults based on graph neural networks according to claim 2, wherein: Performing linear projection on the time series data to obtain projection features; The formula is: , the projection feature formula is: ; Where X represents the input industrial equipment time series data, It is the projection feature obtained by linear projection of X, and its dimension is , is the weight matrix in the linear projection operation, is the bias vector in the linear projection operation, Indicates the number of output channels of the convolutional neural network; Performing convolution processing on the projection feature through a first layer of one-dimensional convolution to obtain a first convolution output; The first convolution output is normalized and activated to obtain the first time feature. Assuming the convolution kernel size of the first layer CNN is 1, the output is ,in , after normalization and ReLU activation function processing; The first time feature is convolved by the second layer of one-dimensional convolution and a residual connection is added to obtain the time feature. The formula is: , the final feature The temporal structure information of the input is preserved and mapped to the convolutional feature dimension.

4. The method for detecting reconstruction error faults based on graph neural networks according to claim 1, wherein: The step of constructing a graph structure based on the time series data, wherein the graph structure represents spatial associations between nodes, includes: For each time step in the time series data, calculate the neighborhood distance of each node feature in a preset space; Determine the adjacency relationship based on the preset radius threshold, generate an adjacency matrix, and generate an edge set based on the adjacency matrix , the edge set is used to construct the graph structure, the formula is: ； generating an edge set according to the adjacency matrix, wherein the edge set is used to construct the graph structure; in, It represents the distance between the feature vectors of node i and node j, r is the preset radius threshold, which is a pre-set constant.

5. The method for detecting reconstruction error faults based on graph neural networks according to claim 1, wherein: Determining a fault detection result according to the reconstructed output and the time series data includes: Calculating a reconstruction error between the reconstructed output and the time series data; If the reconstruction error exceeds a preset threshold, it is determined that the corresponding node has a fault; The overall fault status of the industrial equipment is determined based on the fault judgment results of each node.

6. The method for detecting reconstruction error faults based on graph neural networks according to claim 1, wherein: The acquiring of time series data of industrial equipment, wherein the time series data includes features of multiple nodes, includes: Acquire raw data from multiple sensors in industrial equipment; Preprocessing the raw data to obtain time series data including a time window and the number of nodes; Feature extraction is performed on the time series data to obtain a feature vector for each node.

7. The method for detecting reconstruction error faults based on graph neural networks according to any one of claims 1 to 6, characterized in that: The method comprehensively models the temporal variation trend and spatial correlation of measurement point data by combining convolutional neural networks and graph neural networks.

8. The method for detecting reconstruction error faults based on graph neural networks according to claim 1, wherein: The spatial relationship between measurement points is automatically constructed by generating a radius-based dynamic neighborhood graph structure to adapt to different equipment and measurement point configurations. The algorithm's robustness to noisy and nonlinear data is enhanced through layer-by-layer feature extraction and residual fusion of multi-layer CNN and GATv2.

Images